Liang - 2013 - Formulation For Scalable Optimization of Microcavities Via The

Formulation for scalable optimization of microcavities
via the frequency-averaged local density of states
The MIT Faculty has made this article openly available. Please share
how this access benefits you. Your story matters.
Citation Liang, Xiangdong, and Steven G. Johnson. “Formulation for scalable

optimization of microcavities via the frequency-averaged local
density of states.” Optics Express 21, no. 25 (December 6, 2013):
30812. © 2013 Optical Society of America.
As Published http://dx.doi.org/10.1364/OE.21.030812
Publisher Optical Society of America
Version Final published version
Citable link http://hdl.handle.net/1721.1/82939
Terms of Use Article is made available in accordance with the publisher's

policy and may be subject to US copyright law. Please refer to the
publisher's site for terms of use.
Formulation for scalable optimization of
microcavities via the frequency-averaged
local density of states
Xiangdong Liang∗ and Steven G. Johnson

Department of Mathematics, Massachusetts Institute of Technology, Cambridge, MA 02139,
USA
∗ xdliang@math.mit.edu
Abstract: We present a technique for large-scale optimization of optical

microcavities based on the frequency-averaged local density of states
(LDOS), which circumvents computational difficulties posed by previous
eigenproblem-based formulations and allows us to perform full topology
optimization of three-dimensional (3d) leaky cavity modes. We present
theoretical results for both 2d and fully 3d computations in which every
pixel of the design pattern is a degree of freedom (“topology optimization”),
e.g. for lithographic patterning of dielectric slabs in 3d. More importantly,
we argue that such optimization techniques can be applied to design
cavities for which (unlike silicon-slab single-mode cavities) hand designs
are difficult or unavailable, and in particular we design minimal-volume
multi-mode cavities (e.g. for nonlinear frequency-conversion applications).
© 2013 Optical Society of America
OCIS codes: (120.4570) Optical design of instruments; (140.3945) Microcavities; (230.5750)
Resonators.
References and links

1. M. Nomura, K. Tanabe, S. Iwamoto, and Y. Arakawa, “High-Q design of semiconductor-based ultrasmall pho-
tonic crystal nanocavity,” Opt. Express 18, 8144–8150 (2010).
2. B. Bourdin, “Filters in topology optimization,” Int. J. Numer. Methods Eng. 50, 2143–2158 (2001).
3. M. P. Bendsoe and O. Sigmund, Topology Optimization: Theory, Methods and Applications, 2nd ed. (Springer,
2003).
4. M. Y. Wang, X. Wang, and D. Guo, “A level set method for structural topology optimization,” Comput. Methods
Appl. Mech. Eng. 192, 227–246 (2003).
5. G. Allaire, F. Jouve, and A.-M. Toader, “Structural optimization using sensitivity analysis and a level-set method,”
J. Comput. Phys. 194, 363–393 (2004).
6. O. Sigmund, “Manufacturing tolerant topology optimization,” Acta Mech. Sinica 25, 227–239 (2009).
7. E. Andreassen, A. Clausen, M. Schevenels, B. Lazarov, and O. Sigmund, “Efficient topology optimization in
MATLAB using 88 lines of code,” Struct. Multidiscip. Optim. 43, 1–16 (2011).
8. J. S. Jensen and O. Sigmund, “Topology optimization for nano-photonics,” Laser Photonics Rev. 5, 308–321
(2011).
9. J. Bravo-Abad, A. Rodriguez, P. Bermel, S. G. Johnson, J. D. Joannopoulos, and M. Soljacic, “Enhanced nonlin-
ear optics in photonic-crystal microcavities,” Opt. Express 15, 16161–16176 (2007).
10. R. W. Boyd, Nonlinear Optics, 3rd ed. (Academic, 2008).
11. H. Hashemi, C. W. Qiu, A. P. McCauley, J. D. Joannopoulos, and S. G. Johnson, “A diameter–bandwidth product
limitation of isolated-object cloaking,” Phys. Rev. A 86, 013804 (2012).
12. D. P. Bertsekas, Nonlinear Programming, 2nd ed. (Athena Scientific, 1999).
13. B. E. A. Saleh and M. C. Teich, Fundamentals of Photonics (Wiley Series in Pure and Applied Optics), 2nd ed.
(Wiley-Interscience, 2007).
14. J. D. Joannopoulos, S. G. Johnson, J. N. Winn, and R. D. Meade, Photonic Crystals: Molding the Flow of Light,
2nd ed. (Princeton University, 2008).
#197718 - $15.00 USD Received 16 Sep 2013; revised 17 Nov 2013; accepted 19 Nov 2013; published 6 Dec 2013
(C) 2013 OSA 16 December 2013 | Vol. 21, No. 25 | DOI:10.1364/OE.21.030812 | OPTICS EXPRESS 30812
15. D. G. Rabus, Integrated Ring Resonators, Vol. 127 of Springer Series in Optical Sciences (Springer,2007).
16. S. Lin, E. Schonbrun, and K. Crozier, “Optical manipulation with planar silicon microring resonators,” Nano
Lett. 10, 2408–2411 (2010).
17. J. Vuckovic, M. Loncar, H. Mabuchi, and A. Scherer, “Optimization of the Q factor in photonic crystal micro-
cavities,” IEEE J. Quantum Electron. 38, 850–856 (2002).
18. C. W. Wong, P. T. Rakich, S. G. Johnson, M. Qi, H. I. Smith, E. P. Ippen, L. C. Kimerling, Y. Jeon, G. Barbastathis,
and S. G. Kim, “Strain-tunable silicon photonic band gap microcavities in optical waveguides,” Appl. Phys. Lett.
84, 1242–1244 (2004).
19. Y. Akahane, T. Asano, B.-S. Song, and S. Noda, “High-Q photonic nanocavity in a two-dimensional photonic
crystal,” Nature 425, 944–947 (2003).
20. Y. Akahane, T. Asano, B.-S. Song, and S. Noda, “Fine-tuned high-Q photonic-crystal nanocavity,” Opt. Express
13, 1202–1214 (2005).
21. B.-S. Song, S. Noda, T. Asano, and Y. Akahane, “Ultra-high-Q photonic double-heterostructure nanocavity,”
Nat. Mater. 4, 207–210 (2005).
22. D. C. Dobson and F. Santosa, “Optimal Localization of Eigenfunctions in an Inhomogeneous Medium,” SIAM
J. Appl. Math. 64, 762–774 (2004).
23. C.-Y. Kao and F. Santosa, “Maximization of the quality factor of an optical resonator,” Wave Motion 45, 412–427
(2008).
24. W. R. Frei, H. T. Johnson, and K. D. Choquette, “Optimization of a single defect photonic crystal laser cavity,”
J. Appl. Phys. 103, 033102 (2008).
25. J. Lu and J. Vuckovic, “Inverse design of nanophotonic structures using complementary convex optimization,”
Opt. Express 18, 3793–3804 (2010).
26. A. W. Snyder and J. Love, Optical Waveguide Theory, Science Paperbacks, 190 (Springer, 1983).
27. R. Coccioli, M. Boroditsky, K. W. Kim, Y. Rahmat-Samii, and E. Yablonovitch, “Smallest possible electromag-
netic mode volume in a dielectric cavity,” IEE Proceedings - Optoelectronics 145, 391–397 (1998).
28. E. M. Purcell, “Spontaneous emission probabilities at radio frequencies,” Phys. Rev. 69, 674 (1946).
29. L. Novotny and B. Hecht, Principles of Nano-Optics (Cambridge University, 2006).
30. A. F. Koenderink, “On the use of Purcell factors for plasmon antennas,” Opt. Lett. 35, 4208–4210 (2010).
31. B. Zhen, S.-L. Chua, J. Lee, A. W. Rodriguez, X. Liang, S. G. Johnson, J. D. Joannopoulos, M. Soljačić, and
O. Shapira, “Enabling enhanced emission and low-threshold lasing of organic molecules using special Fano
resonances of macroscopic photonic crystals,” Proc. Natl. Acad. Sci. U. S. A. 110, 13711–13716 (2013).
32. E. A. J. Marcatili, “Bends in optical dielectric guides,” Bell Syst. Tech. J. 48, 2103–2132 (1969).
33. F. Ladouceur, “Roughness, inhomogeneity, and integrated optics,” J. Lightwave Technol. 15, 1020–1025 (1997).
34. V. S. Ilchenko, P. S. Volikov, V. L. Velichansky, F. Treussart, V. Lefèvre-Seguin, J. M. Raimond, and S. Haroche,
“Strain-tunable high-Q optical microsphere resonator,” Opt. Commun. 145, 86–90 (1998).
35. M. Soltani, S. Yegnanarayanan, and A. Adibi, “Ultra-high Q planar silicon microdisk resonators for chip-scale
silicon photonics,” Opt. Express 15, 4694–4704 (2007).
36. L. N. Trefethen and D. Bau, Numerical Linear Algebra (SIAM, 1997).
37. K. Inoue and K. Ohtaka, Photonic Crystals: Physics, Fabrication and Applications, Springer Series in Optical
Sciences (Springer, 2010).
38. A. Oskooi and S. G. Johnson, “Electromagnetic wave source conditions,” in Advances in FDTD Computational
Electrodynamics: Photonics and Nanotechnology, A. Taflove, A. Oskooi, and S. G. Johnson, eds. (Artech, 2013),
Chap. 4, pp. 65–100.
39. J.-M. Gerard and B. Gayral, “Strong Purcell effect for InAs quantum boxes in three-dimensional solid-state
microcavities,” J. Lightwave Technol. 17, 2089–2095 (1999).
40. O. J. F. Martin and N. B. Piller, “Electromagnetic scattering in polarizable backgrounds,” Phys. Rev. E 58, 3909–
3915 (1998).
41. G. D’Aguanno, N. Mattiucci, M. Centini, M. Scalora, and M. J. Bloemer, “Electromagnetic density of modes for
a finite-size three-dimensional structure,” Phys. Rev. E 69, 057601 (2004).
42. J. D. Jackson, Classical Electrodynamics, 2nd ed. (John Wiley, 1975).
43. W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, Numerical Recipes: The Art of Scientific
Computing, 3rd ed. (Cambridge University, 2007).
44. M. E. Peskin and D. V. Schroeder, An Introduction To Quantum Field Theory (Frontiers in Physics) (Westview,
1995).
45. S. G. Johnson, “Numerical methods for computing Casimir interactions,” in Casimir Physics, Vol. 834 of Lecture
Notes in Physics, D. Dalvit, P. Milonni, D. Roberts, and F. da Rosa, eds. (Springer, 2011), Chap. 6.
46. L. D. Landau, L. P. Pitaevskii, and E. M. Lifshitz, Electrodynamics of Continuous Media, 2nd ed. (Butterworth-
Heinemann, 1984).
47. W. C. Chew, Waves and Fields in Inhomogeneous Media (IEEE, 1995).
48. N. A. P. Nicorovici, R. C. McPhedran, and L. C. Botten, “Relative local density of states for homogeneous lossy
materials,” Physica B 405, 2915–2919 (2010).
49. L. V. Ahlfors, Complex Analysis, 3rd ed. (McGraw-Hill, 1978).
50. A. W. Rodriguez, A. P. McCauley, J. D. Joannopoulos, and S. G. Johnson, “Theoretical ingredients of a Casimir
analog computer,” Proc. Natl. Acad. Sci. U. S. A. 107, 9531–9536 (2010).
51. X. Liang, Ph.D. thesis, Massachusetts Institute of Technology, 2013.
52. S. Scheel, L. Knöll, and D. G. Welsch, “Spontaneous decay of an excited atom in an absorbing dielectric,” Phys.
Rev. A 60, 4094–4104 (1999).
53. C. Van Vlack and S. Hughes, “Finite-difference time-domain technique as an efficient tool for calculating the
regularized Green function: applications to the local-field problem in quantum optics for inhomogeneous lossy
materials,” Opt. Lett. 37, 2880–2882 (2012).
54. G. Strang, Computational Science and Engineering (Wellesley-Cambridge, 2007).
55. A. Christ and H. L. Hartnagel, “Three-dimensional finite-difference method for the analysis of microwave-device
embedding,” IEEE Trans. Microwave Theory 35, 688–696 (1987).
56. K. Yasumoto, Electromagnetic Theory and Applications for Photonic Crystals, Optical Science and Engineering
(CRC, 2005).
57. W. Shin and S. Fan, “Choice of the perfectly matched layer boundary condition for frequency-domain Maxwell’s
equations solvers,” J. Comput. Phys. 231, 3406–3431 (2012).
58. T. A. Davis, Direct Methods for Sparse Linear Systems (Fundamentals of Algorithms) (SIAM, 2006).
59. S. Balay, W. D. Gropp, L. C. McInnes, and B. F. Smith, “Efficient Management of Parallelism in Object Oriented
Numerical Software Libraries,” In Modern Software Tools in Scientific Computing, E. Arge, A. M. Bruaset, and
H. P. Langtangen, eds. (Birkhäuser, 1997).
60. S. Balay, J. Brown, K. Buschelman, V. Eijkhout, W. D. Gropp, D. Kaushik, M. G. Knepley, L. C. McInnes,
B. F. Smith, and H. Zhang, “PETSc Users Manual,” Technical Report No. ANL-95/11 - Revision 3.3, Argonne
National Laboratory (2012).
61. S. Balay, J. Brown, K. Buschelman, W. D. Gropp, D. Kaushik, M. G. Knepley, L. C. McInnes, B. F. Smith, and
H. Zhang, “PETSc Web page,” http://www.mcs.anl.gov/petsc, 2012.
62. P. Hénon, P. Ramet, and J. Roman, “PaStiX: a high-performance parallel direct solver for sparse symmetric
positive definite systems,” Parallel Computing 28, 301–321 (2002).
63. S. G. Johnson, “The NLopt nonlinear-optimization package,” http://ab-initio.mit.edu/nlopt.
64. D. Liu and J. Nocedal, “On the limited memory BFGS method for large scale optimization,” Math. Program. 45,
503–528 (1989).
65. K. Svanberg, “A class of globally convergent optimization methods based on conservative convex separable
approximations,” SIAM J. Optimiz. 12, 555–573 (2002).
66. J. Nocedal and S. J. Wright, Numerical Optimization (Springer, 2000).
67. A. Mutapcic, S. Boyd, A. Farjadpour, S. G. Johnson, and Y. Avniel, “Robust design of slow-light tapers in
periodic waveguides,” Eng. Optimiz. 41, 365–384 (2009).
68. A. Oskooi, A. Mutapcic, S. Noda, J. D. Joannopoulos, S. P. Boyd, and S. G. Johnson, “Robust optimization of
adiabatic tapers for coupling to slow-light photonic-crystal waveguides,” Opt. Express 20, 21558–21575 (2012).
69. Y. Xu, W. Liang, A. Yariv, J. G. Fleming, and S.-Y. Lin, “High-quality-factor Bragg onion resonators with omni-
directional reflector cladding,” Opt. Lett. 28, 2144–2146 (2003).
70. Z. Artstein, “Discrete and continuous bang-bang and facial spaces or: Look for the extreme points,” SIAM Rev.
22, 172–185 (1980).
71. A. F. Oskooi, Ph.D. thesis, Massachusetts Institute of Technology, 2010.
72. J. Lu, S. Boyd, and J. Vuckovic, “Inverse design of a three-dimensional nanophotonic resonator,” Opt. Express
19, 10563–10570 (2011).
73. B. Osting and M. I. Weinstein, “Long-lived scattering resonances and Bragg structures,” SIAM J. Appl. Math.
73, 827–852 (2013).
74. T. Inui, Y. Tanabe, and Y. Onodera, Group Theory and Its Applications in Physics (Springer, 1996).
75. S. Boyd and L. Vandenberghe, Convex Optimization (Cambridge University, 2004).
76. P. D. Drummond, K. J. McNeil, and D. F. Walls, “Non-equilibrium transitions in sub/second harmonic genera-
tion,” Opt. Acta 27, 321–335 (1980).
77. L.-A. Wu, M. Xiao, and H. J. Kimble, “Squeezed states of light from an optical parametric oscillator,” J. Opt.
Soc. Am. B 4, 1465–1475 (1987).
78. Z. Y. Ou and H. J. Kimble, “Enhanced conversion efficiency for harmonic generation with double resonance,”
Opt. Lett. 18, 1053–1055 (1993).
79. A. Rodriguez, M. Soljacic, J. D. Joannopoulos, and S. G. Johnson, “χ (2) and χ (3) harmonic generation at a
critical power in inhomogeneous doubly resonant cavities,” Opt. Express 15, 7303–7318 (2007).
80. I. B. Burgess, Y. Zhang, M. W. McCutcheon, A. W. Rodriguez, J. Bravo-Abad, S. G. Johnson, and M. Loncar,
“Design of an efficient terahertz source using triply resonant nonlinear photonic crystal cavities,” Opt. Express
17, 20099–20108 (2009).
81. Z.-F. Bi, A. W. Rodriguez, H. Hashemi, D. Duchesne, M. Loncar, K.-M. Wang, and S. G. Johnson, “High-
efficiency second-harmonic generation in doubly-resonant χ (2) microring resonators,” Opt. Express 20, 7526–
7543 (2012).
82. M. W. McCutcheon and M. Loncar, “Design of a silicon nitride photonic crystal nanocavity with a Quality factor
of one million for coupling to a diamond nanocrystal,” Opt. Express 16, 19136–19145 (2008).
83. A. R. Md Zain, N. P. Johnson, M. Sorel, and R. M. De La Rue, “Ultra high quality factor one dimensional pho-
tonic crystal/photonic wire micro-cavities in silicon-on-insulator (SOI),” Opt. Express 16, 12084–12089 (2008).
84. Z. M. Meng, F. Qin, Y. Liu, and Z. Y. Li, “High-Q microcavities in low-index one-dimensional photonic crystal
slabs based on modal gap confinement,” J. Appl. Phys. 109, 043107 (2011).
85. J. T. Robinson, C. Manolatou, L. Chen, and M. Lipson, “Ultrasmall mode volumes in dielectric optical micro-
cavities,” Phys. Rev. Lett. 95, 143901 (2005).
86. M. Nomura, “GaAs-based air-slot photonic crystal nanocavity for optomechanical oscillators,” Opt. Express 20,
5204–5212 (2012).
87. S. Kita, K. Nozaki, S. Hachuda, H. Watanabe, Y. Saito, S. Otsuka, T. Nakada, Y. Arita, and T. Baba, “Photonic
Crystal Point-Shift Nanolasers With and Without Nanoslots Design, Fabrication, Lasing, and Sensing Character-
istics,” IEEE J. Sel. Top. Quantum Electron. 17, 1632–1647 (2011).
88. Y. Li, J. Zheng, J. Gao, J. Shu, M. S. Aras, and C. W. Wong, “Design of dispersive optomechanical coupling and
cooling in ultrahigh-Q/V slot-type photonic crystal cavities,” Opt. Express 18, 23844–23856 (2010).
89. O. Sigmund and J. Petersson, “Numerical instabilities in topology optimization: A survey on procedures dealing
with checkerboards, mesh-dependencies and local minima,” Struct. Multidiscip. Optim. 16, 68–75 (1998).
90. S. J. Osher and R. P. Fedkiw, Level Set Methods and Dynamic Implicit Surfaces (Springer, 2002).
91. C. Y. Kao, S. Osher, and E. Yablonovitch, “Maximizing band gaps in two-dimensional photonic crystals by using
level set methods, Appl. Phys. B 81, 235–244 (2005).
92. J. K. Guest, J. H. Prévost, and T. Belytschko, “Achieving minimum length scale in topology optimization using
nodal design variables and projection functions,” Int. J. Numer. Methods Eng. 61, 238–254 (2004).
93. O. Sigmund, “Morphology-based black and white filters for topology optimization,” Struct. Multidiscip. Optim.
33, 401–424 (2007).
94. S. Xu, Y. Cai, and G. Cheng, “Volume preserving nonlinear density filter based on heaviside functions,” Struct.
Multidiscip. Optim. 41, 495–505 (2010).
95. F. Wang, B. Lazarov, and O. Sigmund, “On projection methods, convergence and robust formulations in topology
optimization,” Struct. Multidiscip. Optim. 43, 767–784 (2011).
96. A. Ben-Tal, L. El Ghaoui, and A. S. Nemirovski, Robust Optimization (Princeton University, 2009).
97. D. Bertsimas, O. Nohadani, and K. M. Teo, “Robust optimization for unconstrained simulation-based problems,”
Oper. Res. 58, 161–178 (2010).
98. F. Wang, J. S. Jensen, and O. Sigmund, “Robust topology optimization of photonic crystal waveguides with
tailored dispersion properties,” J. Opt. Soc. Am. B 28, 387–397 (2011).
99. H. Men, R. M. Freund, N. C. Nguyen, J. Saa-Seoane, and J. Peraire, “Fabrication-adaptive optimization with an
application to photonic crystal design,” arXiv:1307.5571.
100. A. Faraon, C. Santori, Z. Huang, V. M. Acosta, and R. G. Beausoleil, “Coupling of nitrogen-vacancy centers to
photonic crystal cavities in monocrystalline diamond,” Phys. Rev. Lett. 109, 033604 (2012).
101. L. Li, M. Trusheim, O. Gaathon, K. Kisslinger, C.-J. Cheng, M. Lu, D. Su, X. Yao, H.-C. Huang, I. Bayn, A.
Wolcott, R. M. Osgood Jr., and D. Englund, “Reactive ion etching: Optimized diamond membrane fabrication
for transmission electron microscopy,” J. Vac. Sci. Technol. B 31, 06FF01 (2013).
102. Z. Yu, A. Raman, and S. Fan, “Fundamental limit of light trapping in grating structures,” Opt. Express 18, A366–
A380 (2010).
103. Z. Yu, A. Raman, and S. Fan, “Fundamental limit of nanophotonic light trapping in solar cells,” Proc. Natl. Acad.
Sci. U. S. A. 107, 17491–17496 (2010).
1. Introduction
In this paper, we present a new technique for large-scale optimization of optical microcavi-
ties based on the frequency-averaged local density of states (LDOS), which circumvents com-
putational difficulties posed by previous eigenproblem-based formulations and allows us to
perform full topology optimization of three-dimensional (3d) leaky cavity modes. Essentially,
this technique allows us to minimize modal volume V for a given cavity quality factor Q, or
equivalently a given bandwidth, by solving a single complex-frequency scattering problem at
each optimization step, which is both computationally easier than solving an eigenproblem and
avoids the difficulty of selecting which eigenvalue to optimize. We present proof-of-concept
results in both 2d and 3d computations in which every pixel of the design pattern is a degree
of freedom (“topology optimization”), e.g. for lithographic patterning of dielectric slabs in 3d.
For a design Q of ≈ 104 in 3d for silicon slabs (index n = 3.52), we obtain a modal vol-
ume V = 0.06(λ /n)3 (relative to the vacuum wavelength λ ), which is 4× smaller than the best
comparable-Q volume found in the literature [1]. Here, the focus is on illustrating the technique
rather than on designing practical cavities (for which many good designs are already available
in the single-mode silicon case), so we do not incorporate regularizations [2–8] to force the
optimization to find easily fabricated structures. However, our results are still useful in estab-
lishing theoretical bounds, showing that significant room for improvement remains even for
silicon cavities. More importantly, we show that such optimization techniques can be applied
to design cavities for which (unlike silicon-slab single-mode cavities) hand designs are difficult
or unavailable, and in particular we design proof-of-concept minimal-volume multi-mode cav-
ities (e.g. for nonlinear frequency-conversion applications [9, 10]) in Secs. 8.2.3 and 8.3. We
find that the optimum doubly degenerate cavity appears to be three-fold symmetric, while two-
frequency cavities have more complex shapes depending on the polarization, although further
work is needed to obtain manufacturable structures. Theoretical optimization is also useful to
investigate the tradeoff between the size of the design region and the maximum attainable Q
in slab-like situations (Sec. 8.4), and find it to be roughly exponential. All of these results are
enabled by a sequence of mathematical transformations of the original cavity-design problem,
depicted in Fig. 1: from the Purcell factor Q/V to the more well-posed problem of minimizing
V for a given Q (Sec. 2), from the eigenproblem to a scattering problem via the LDOS (Sec. 3),
from a frequency-averaged LDOS to a single complex-frequency scattering problem [11] via
contour integration (Secs. 4–5), and finally from maximizing LDOS to minimizing 1/LDOS
in order to circumvent optimization problems that arise for sharply peaked objectives (Sec. 7).
Even though the optimization problem is nonlinear and probably non-convex [12], we obtain
similar optima for many initial structures (including vacuum or random pixels), suggesting that
the result may be close to a global optimum.
Fig. 1. Starting with the naive objective of maximizing a microcavity’s Purcell factor Q/V ,
we perform a sequence of transformations of the problem in order to make it well posed
and tractable. Here, we give a schematic diagram of each transformation, along with the
corresponding section of the paper in which they are discussed.
Microcavity design, which seeks to confine a cavity “mode” for a long time (dimensionless
lifetime Q) in a small “modal volume” V , has a long history [13, 14], and until recently has
been dominated by hand designs, from ring resonators [15, 16] to photonic-crystal slab cavi-
ties [17, 18], in which the parameters are manually tweaked to obtain the desired results (or in
some cases computer optimization is performed over a handful of parameters [19–21]). These
manual designs have been tremendously successful, attaining experimentally verified Q of 106
or more in modal volumes only a few wavelengths in diameter [19–21]. However, fundamental
questions remain: how close are the existing designs to the theoretical optima (e.g. the minimal
V for any given Q), and do the optimal designs resemble the hand designs or are they entirely
different? (As discussed in Sec. 2, the question is not what is the maximum Q or Q/V , because
those questions have a trivial theoretical answer: ∞.) Furthermore, while design of single-mode
cavities in silicon slabs has been heavily studied and many design heuristics are known, very
different designs and grueling effort may be required in radically different circumstances, from
new materials to new design goals such as multi-mode cavities for nonlinear optics [10]. Large-
scale “topology” optimization, in which the design pattern is completely determined by compu-
tational search (often with thousands of free parameters) with little or no a priori information,
offers a different route to addressing these questions. Although several authors have developed
topology-optimization techniques for microcavity design [22–25], most of that previous work
was limited to 2d calculations [22, 23] and none of the previous work fully addressed the de-
sign tradeoff between Q and V . Some topology optimization dealt with lossless systems and
avoided Q entirely [22], or maximized a 2d Q without controlling V [23] (which we argue in
Sec. 2 leads to an ill-posed optimization problem). Other work used a “2.5d” heuristic for the
cavity radiation loss [25], which permitted 2d calculations and led to intriguing designs, but
limits the attainable Q (to < 104) because there appears to be no systematic way to improve
this heuristic loss estimate. One group did perform fully 3d calculations with absorbing bound-
aries to capture radiation loss [24], but limited their degrees of freedom to a small region inside
a photonic crystal and included an unphysical absorbing material within the cavity itself, lead-
ing to a heuristic objective whose calculation does not appear to have a rigorous quantitative
relationship to key cavity properties. As we review in Sec. 3, maximizing LDOS corresponds
mathematically to maximizing the power radiated by a dipole current source inside the cavity.
Naively, this may seem backwards: isn’t the purpose of a cavity to minimize radiated power?
However, that intuition stems from the situation in which the energy density inside the cavity
is held constant: there is an equation Q = ω U/P relating the cavity Q and frequency ω with
the radiated power P and the confined energy U [14], so that P ∼ 1/Q for fixed U. In our case,
however, we are holding the current amplitude in the cavity fixed, in which case the radiated
power P ∼ Q (as reviewed in Sec. 3) while one can show from a coupled-mode framework [14]
that U ∼ Q2 . Conversely, minimizing radiated power for a fixed current would correspond to
minimizing LDOS, which would result in the absence of resonances. (Interestingly, Frei et
al. [24] actually minimized radiated power from a fixed-current dipole, but their introduction
of a heuristic absorbing region inside the cavity apparently compensated for this inverted ob-
jective and resulted in a resonant mode, albeit a mode optimizing an unclear objective.) On the
other hand, our work adapts two crucial ideas from [24]: solving a scattering problem rather
than an eigenproblem, and optimizing over a finite bandwidth. Nor did any previous work, to
our knowledge, address topology optimization of multi-mode cavities.
2. Eigenproblem formulation
There are two key figures of merit for a resonant mode En (x) of a cavity: quality factor Q and
modal volume V . The quality factor Q is a dimensionless lifetime, and 1/Q is a dimensionless
decay rate [14]. Mathematically, Q is related to the frequency-domain Maxwell eigenvalue
problem:
1
∇× ∇ × En (x) = ωn2 ε (x)En (x) (1)
µ (x)
with radiation boundary conditions. Because of the lossy boundary conditions, the eigenprob-
lem is non-Hermitian and the eigenvalues are complex. The Q for the mode En (x) with eigen-
frequency ωn [14] is
Re[ωn ]
Q= . (2)
−2 Im[ωn ]
Technically, if one considers an infinite open system (as opposed to a finite system with
some absorbing boundary layer), a number of mathematical subtleties arise because the leaky
“modes” are not true normalizable eigenfunctions, but are rather related to the residues of poles
in the Green’s function [26]; however, we will circumvent all of these difficulties in this paper
by transition to a Green’s-function LDOS approach in Sec. 3 that does not deal explicitly with
eigenvalue problems. The modal volume V [27], defined as
ε (x)|En (x)|2 dx
R
V= , (3)
max{ε (x)|En (x)|2 }
is a measure of the volume within which the mode is confined. (Although Eq. (3) is a standard
expression, it should really be viewed as the modal volume in the limit of a lossless cavity. For
a lossy cavity, ε (x)|En (x)|2 dx does not converge in an infinite open system; a more rigorous
R
treatment is to define the LDOS in terms of the Green’s function as in Sec. 3 and obtain the
Purcell factor from the ratio with the LDOS of the homogeneous medium.) The Purcell factor
[28], which describes the enhancement of spontaneous emission rate, can be written as [29, 30]
λ
3
3Q
. (4)
4π 2V n
Here λ is the vacuum wavelength and n is the index of refraction.

For applications with light-matter interactions (such as lasers, sensors, and nonlinear fre-
quency converters), maximal lifetime Q and minimal modal volume V are desirable [31]. It is
therefore tempting to use the Purcell factor in Eq. (4) or Q/V as the figure of merit for cavity
optimization. Unfortunately, maximizing Q/V leads to an ill-posed problem, because the maxi-
mum of Q/V is ∞; for example, Q/V grows exponentially with radius for a ring resonator [32].
In practice, any optimization in a finite computation cell will obtain a finite Q and V [23], but
the values are then just an artifact of the finite computational domain; in this sense, maximizing
Q/V is not well-posed because the solution does not converge as one increases the size of the
computational domain.
In practice, however, there is an upper bound on the useful Q for two reasons. First, besides
the intrinsic radiation loss (Qrad ) in a cavity, there are also radiation losses due to surface rough-
ness (Qroughness ) and material absorption (Qabsorption ). The total loss rate 1/Qtotal is the sum of
these three effects [14]:
1 1 1 1
= + + . (5)
Qtotal Qrad Qroughness Qabsorption
In real applications, therefore, the Qtotal cannot be arbitrarily large. For example, in integrated
optics it is difficult to get Qtotal more than a few million due to surface roughness [33], although
microdisk and microsphere resonators can achieve higher Q by delocalizing the mode away
from the surface [34, 35]. Second, there is another quality factor in the system. For any cavity-
based device, the cavity is always intentionally coupled to some channels (e.g., waveguides)
to get light in and out. That coupling process will be described by its own lifetime Qcoupling . It
turns out that the losses in such a coupled device are proportional to Qcoupling /Qtotal [14]. Once
these losses are decreased below the desired loss budget, it often does not matter in practice if
one decreases them further.
A better optimization problem might be, instead, to maximize Q/V subject to Q = Q̃, or
alternatively:
minV
(6)
s.t. Q ≥ Q̃,
where Q̃ is determined by the bandwidth and loss tolerance of applications. By solving the
non-Hermitian Maxwell eigenproblem Eq. (1), one could obtain Q and V from eigenvectors
En (x) and eigenvalues ωn through Eq. (2) and Eq. (3). Then a natural question to ask is which
eigenvalue one should optimize. In practice, one has some design frequency ω̃ given by the
application, so one could optimize the eigenvalue closest to ω̃ . (Equivalently, thanks to the
scale-invariance of Maxwell’s equations [14], we choose units so that ω̃ = 2π , in which case
the main computational choice is the resolution, i.e. the number of pixels per wavelength λ ,
and in some cases a slab thickness/λ .) However, asking for the mode closest to ω̃ leads to
discontinuities: as the structure changes during optimization, the mode closest to ω̃ will tend to
hop discontinuously. Although there are some ways to deal with this [22], the problem becomes
worse when one simulates the radiation loss in a patterned dielectric slab, because in this case
the finite cell is approximating a continuum of radiation modes above and below the slab. As
a result, there are more and more closely spaced modes as the cell size increases. Hence, we
want to circumvent this difficulty by adopting a new approach: turning the eigenproblem into
a linear scattering problem. This also reduces the computational expense, because we will now
require only a single linear solve per structure, whereas finding eigenvalues in the interior of a
spectrum (e.g. by the shift-and-invert algorithm) requires many linear solves [36].
3. LDOS formulation
The well-known Purcell factor (Q/V and variations thereof) is, in fact, only an approximation of
a more fundamental quantity, the local density of states (LDOS), which is defined in terms of the
Green’s function of the system (and can be related to a “density of modes” per unit frequency
per unit volume) [37]. Paradoxically, the exact LDOS is easier to compute than Q/V , because
the former involves only a scattering problem (a linear system of equations) whereas the latter
involves a non-Hermitian eigenproblem (for a leaky mode in the interior of the spectrum). In
this section, we briefly review the definition of the LDOS and how it relates to Q/V , and in the
next section we describe how the LDOS can be used as the objective for a well-posed cavity
optimization problem.
In particular, we begin with the per-polarization LDOS (partial LDOS), denoted by
LDOS j (ω , x′ ) for a polarization in direction j, which is proportional to the power radiated
by a dipole in direction (unit vector) ê j at position x′ with a frequency ω , i.e. a current
J ∼ ê j e−iω t δ (x − x′ ). (The total LDOS is simply ∑ j LDOS j , proportional to power radiated
by a randomly oriented dipole.) This is a key physical quantity because quantum fluctuations,
e.g. spontaneous emission or thermal fluctuations, can be viewed semiclassically as dipole cur-
rents, and the LDOS yields the enhancement of these fluctuations by any given geometry. For
example, the spontaneous emission rate is proportional to the LDOS [29, 38], and therefore
the ratio of LDOS between two systems indicates enhancement or suppression of spontaneous
emission. The Purcell factor Q/V is an approximation to this LDOS enhancement for a high-Q
microcavity, assuming that the emitter (e.g. atom or quantum dot) is positioned and oriented at
the location of peak LDOS j in the cavity [29, 30, 38, 39].
These quantities can be defined more precisely as follows [29, 37, 38, 40, 41]. Poynting’s
theorem [42] implies that the power radiated by a dipole J(x) is
1
Z
Pj (ω , x′ ) = − Re [J∗ (x) · E(x)]dx, (7)
2
where E(x) is the total electric field solving the frequency-domain scattering problem
M (ε , ω )E(x) = iω J(x)
1
M (ε , ω ) = ∇ × ∇ × − ε (x)ω 2 (8)
µ (x)
J(x) = δ (x − x′)ê j .
This yields the following well known [29, 37, 38, 40, 41] definition of LDOS j :
Z
12 6
LDOS j (ω , x ) = Pj (ω , x ) = − Re
′ ′ ∗
J (x) · E(x)dx , (9)
π π
where the 12/π factor is a conventional normalization that arises from the dual interpretation
of the LDOS as a local “density” of eigenstates [29, 38]. (The normalization is irrelevant for
optimization or for evaluating LDOS ratios in different systems to obtain enhancement factors.)
In the limit of a low-loss cavity, in which the LDOS is dominated by the contribution of a
single pole in Maxwell’s equations (a single “resonant mode”), one can derive the approxima-
tion [38]
6 Q
LDOS j (ω , x′ ) = , (10)
πωε (x ) V
′
and by √ taking the ratio of this quantity with the LDOS of a homogeneous medium with index
n = ε one [29] obtains the traditional Purcell factor Eq. (4). In particular, the LDOS in a
microcavity resonating a frequency ω̃ (1 + i/2Q̃) is approximately in the form of a Lorentzian
peak centered at ω̃ with a bandwidth ω̃ /2Q̃ [38], and the Purcell factor is the enhancement at
the peak. This relationship between Q and LDOS bandwidth is the key to obtaining a tractable
well-posed optimization problem in the next section.
We propose that LDOS j or its variants can be used as figure of merit for the characterization
and optimization of a microcavity. Note that the precise figure of merit should really depend
on the application. For example, if we are interested in the spontaneous emission rate for the
dipole at a specific position x′ with a specific polarization ê j , then LDOS j (ω , x′ ) is the most
relevant figure of merit. On the other hand, if we have a dipole at a specific point x′ with
a randomly distributed polarization, then the figure of merit would be ∑ j LDOS j (ω , x′ ) (for
the average case) or min j LDOS j (ω , x′ ) (for the worst case). In nonlinear devices for frequency
conversion (in particular, second harmonic generation), a more relevant figure of merit might be
minn LDOS j (ωn , x′ ), where ωn are different frequencies of interest. Instead of at a single point
x′ , if the dipoles of interest are distributed [with probability density function s(x)] in a region
V with polarization ê j , the most relevant figure of merit in this case is V LDOS j (ω , x)s(x)dx.
R
Depending on the applications for enhancement or inhibition, we should maximize or minimize

the figure of merit correspondingly. Many other variations could be devised.
For all the above mentioned applications, LDOS j is the basic building block. We therefore
focus on this case: we maximize the spontaneous emission rate for a dipole at a point x′ with
given polarization ê j . For simplicity, from now on, we shall omit the explicit j and x′ depen-
dence from LDOS j (ω , x′ ), and simply write LDOS(ω ) to denote the figure of merit given in
Eq. (9). In Sec. 8.2, however, we will also consider a case where the dipole polarization is
randomly distributed, and in Secs. 8.2.3 and 8.3 we consider multi-mode cavities.
4. Frequency-averaged LDOS
In previous section, we proposed that one way of attacking the problem of microcavity design
is to maximize the LDOS. However, simply maximizing LDOS(ω ) in Eq. (9) is still ill-posed
as in Sec 2, being equivalent to Q/V . As explained in Sec. 2, a well-posed formulation could
be obtained by specifying a desired cavity lifetime Q̃. In terms of LDOS, this is equivalent
to specifying an upper bound ω̃ /2Q̃ on the bandwidth Γ̃. For computational convenience, we
instead maximize the average LDOS over this finite bandwidth Γ̃:
Z ∞
L= LDOS(ω )W (ω )dω . (11)
−∞
Here, W (ω ) is some weight function or window function we choose, which is peaked around
the design frequency ω̃ and decays rapidly (with a finite integral) outside of a bandwidth Γ̃
around ω̃ . As we will explain in Sec. 4.3, it will turn out that performing this average is math-
ematically equivalent to computing the LDOS at a single frequency with an absorption loss
added into the system, which effectively limits Q to the desired bandwidth (more precisely, for
Q ≫ Q̃ the figure of merit is dominated by the modal volume). Equivalently, as Q increases the
LDOS of a resonance approaches a delta function with amplitude ∼ 1/V [38], so that the inte-
gral Eq. (11) is determined mainly by the 1/V factor as soon as the resonance is narrower than
Γ̃ (Q ≫ Q̃). The key point is that maximizing the bandwidth-averaged LDOS regularizes the
optimization problem to eliminate the possibility of diverging-Q solutions as the computational
cell size increases.
The main remaining question is how to efficiently compute the average LDOS of Eq. (11).
The most straightforward approach would be to apply a standard numerical-integration pro-
cedure [43], which would involve evaluating the LDOS (i.e., solving a scattering problem) for
many separate frequencies over the bandwidth Γ̃. (This is similar in spirit to the multi-frequency
optimization approach of [24].) However, as described in the next two sections, we can exploit
techniques from complex analysis to evaluate the exact LDOS integral by solving only a sin-
gle scattering problem at a complex frequency ω̃ + iΓ̃. The key to this technique is that the
LDOS derives from a causal Green’s function, as reviewed in Sec. 4.1. This allows us to per-
form a contour integration, with an appropriate choice of W (ω ), in order to obtain the integral
as described in Sec. 4.2. As explained in Sec. 4.3, we can reinterpret any complex-frequency
scattering problem as a real-frequency scattering problem with complex materials (i.e., absorp-
tion). Finally, in Sec 5 we discuss convenient choices of the window function W (ω ) that reduce
the contour integral to a single complex-ω scattering problem. (We previously applied a very
similar application of causality and complex analysis to analysis of electromagnetic cloaking
bandwidth [11], and related ideas can be found in quantum field theory [44, 45].)
4.1. Causality and analyticity

Before we proceed, we first define a function f (ω ), which is a complex version of LDOS(ω ):
6
Z
f (ω , x ) = −
′
J∗ (x) · E(x, ω )dx. (12)
π
Comparing with Eq. (9), it is clear that LDOS(ω ) = Re[ f (ω )]. [From now on, we will omit the
x′ dependency of f (ω , x′ ).] Note that the operator M (ε , ω ) given in Eq. (8) is a linear operator
relating the electric field E(x, ω ) to the (time-harmonic) input electric current J(x) at a given
frequency ω . Causality [the electric field E comes after (not before) the current J] implies that
E(x, ω ) is analytic in the upper-half complex-ω plane [46]. Therefore, f (ω ) is also analytic in
the upper-half complex-ω plane.
Im (ω)
A3 A3
∞
ωk
A2
A1 A1
Re (ω)
Fig. 2. Contour integration path. The frequency-averaged LDOS is the path integral along
arc A1 in the limit of an infinite-radius arc. By choosing the proper window/weight function
W (ω ) for optimizing LDOS in a desired bandwidth, the contribution along arcs A2 and A3
can be made negligible compared to A1 . Therefore, the residues at poles ω̄k enclosed by
this contour can be used to approximate the averaged LDOS.
4.2. Contour integration

In this section, we are going to compute the mean LDOS by exploiting the analyticity of f (ω ),
via:
Z ∞ Z ∞ Z ∞
L= LDOS(ω )W (ω )dω = Re[ f (ω )]W (ω )dω = Re p.v. f (ω )W (ω )dω . (13)
−∞ −∞ −∞
Here, p.v. denotes the Cauchy principal value, which we use because the imaginary part of
f (ω ) may have a singularity at ω = 0, as in the case of f (ω ) in free space [40, 42, 47, 48]. Now
we want to complete our integration contour (Fig. 2) in the upper-half plane and evaluate L by
residue theorem [49]
Z Z Z
+ + f (ω )W (ω )dω = 2π i ∑ Res [ f (ω )W (ω ), ω̄k ] . (14)
A1 A2 A3 k
Here, ω̄k denotes a pole of W (ω ) in the upper-half plane and Res is its residue [49]. In Sec. 5,
we will choose W (ω ) so that these poles and residues are easy to evaluate. Furthermore, we
can choose the weight function W (ω ) so that it decays faster than 1/|ω |3 for large ω , in which
case the contribution from arc A3 will be zero since LDOS(ω ) is proportional to ω 2 in 3d (or
ω in 2d) free space and f (ω )W (ω ) will decay faster than 1/|ω | on arc A3 . We can obtain the
contribution from arc A2 by evaluating the residue due to the simple pole of f (ω ) at ω = 0
1
Z
f (ω )W (ω )dω = − 2π iW (0) lim ω f (ω ). (15)
A2 2 ω →0
The factor −1/2 comes from the fact that the integration is along a clockwise semicircle. Since
the weight function is peaked around design frequency ω̃ with some narrow bandwidth Γ̃, we
can require that W (0) is small, and therefore the contribution from A2 is negligible comparing
to the residues at ω̄k . From Eq. (13), L is just the path integral along the path A1 [46]. Therefore,
we have
" Z Z #
L = Re 2π i ∑ Res [ f (ω )W (ω ), ω̄k ] − + f (ω )W (ω )dω
k A2 A3
" # (16)
≈ Re 2π i ∑ Res [ f (ω )W (ω ), ω̄k ] .
k
4.3. From complex frequency to material absorption

To compute the residue at the complex poles, we need to solve the scattering problem at com-
plex frequencies. More precisely, the scattering problem Eq. (8) at complex frequency ω + iΓ
can be written as

1
∇× ∇ × −ε (x)(ω + iΓ)2 E(x, ω + iΓ) = i(ω + iΓ)J(x)
µ (x)
!
Γ

1
⇐⇒ ∇ × ∇ × −ε (x)ω 1 + i
2
E(x, ω + iΓ) = iω J(x) (17)
µ (x)(1 + i ωΓ ) ω
!
1 i
⇐⇒ ∇ × ∇ × − ε (x) 1 + ω 2 E(x, ω + iΓ) = iω J(x).
µ (x)(1 + 2Q i
) 2Q
We denote this complex scattering operator by M˜(ε , ω ), namely

1 i 1
M˜(ε , ω ) = ∇ × ∇ × − ε (x) 1 + ω2 = ∇ × ∇ × −ε̃ (x)ω 2 . (18)
µ (x)(1 + 2Q
i
) 2Q µ̃ (x)
Clearly, this is equivalent to solving a scattering problem at real frequency ω̃ with materials
ε̃ (x) = ε (x)(1 + i/2Q) and µ̃ (x) = µ (x)(1 + i/2Q). (In fact, any change to the frequency can
be converted into a change of materials [50].) In particular, adding a positive imaginary part
to ω (for ω̄k in the upper-half plane [11]) corresponds to a positive imaginary part in ε̃ (x) and
µ̃ (x), which corresponds (with our e−iω t convention) to an absorption loss.
5. Possible window functions

In this section, we discuss two convenient window functions: a simple Lorentzian and the
square of a Lorentzian. (A third possibility, the difference of two Lorentzians is discussed
in [51].)
5.1. A simple Lorentzian

The simplest window function, with only a simple pole in the upper-half plane, is a Lorentzian
centered at ω̃ with half-width Γ̃. The frequency-average LDOS against this weight is
Γ̃/π
Z
L1 = LDOS(ω ) = Re[ f (ω̃ + iΓ̃)], (19)
(ω − ω̃ )2 + Γ̃2
which only requires solving the scattering problem Eq. (17) at a single complex-frequency
ω̃ + iΓ̃. L1 is a perfectly finite, well-defined quantity in a discretized simulation with a finite
spatial resolution (finite grid).
In combination with Sec. 4.3, the objective Eq. (19) has a simple interpretation that coin-
cides with the discussion in Sec. 2. Our frequency-averaged LDOS objective is equivalent to
maximizing the LDOS at a single frequency, i.e. the Purcell factor Q/V of a cavity, but a cav-
ity in which an absorption loss has been added to otherwise lossless materials. In particular,
the absorption loss that arises from multiplying ε (x) and µ (x) by 1 + i/2Q̃ can be seen from
perturbation theory [14] (for Q̃ ≫ 1) to yield an absorption lifetime of exactly Qabsorption = Q̃.
From Eq. (5), this means that Qtotal ≤ Q̃, and that increasing Qrad ≫ Q̃ will have little effect
on the Purcell factor Qtot /V . Hence, optimizing the frequency-averaged objective will tend to
increase Qrad until it is > Q̃, and after that will mainly try to decrease V , similar in spirit to
Eq. (6).
However, a careful examination reveals that this simple average does not converge as the
resolution increases. There are two equivalent ways to understand this. First, in a continuous
medium, the integral does not converge because the window function decays like 1/|ω |2 while
LDOS(ω ) behaves like |ω | (in 2d free space) or |ω |2 (in 3d free space) for large |ω |. (For finite
spatial resolution, there is an upper frequency cutoff that eliminates this divergence.) Second,
from the relationship between the complex-frequency scattering and lossy material discussed
in the previous section, we know that the residue Re[ f (ω̃ + iΓ̃)] is actually the power emitted
by a dipole in lossy material, which is the sum of the power radiating to the outside of the
cavity and the power absorbed by the lossy material in the cavity [41]. It is known that this
absorbed power is infinite because E(x) diverges as 1/r3 in the neighborhood of the dipole
(in 3d free space) [40, 47, 52, 53]. (In a lossless medium, only Im[E(x)] diverges as 1/r3 , so
LDOS ∼ Re[E(x)] is finite.) In discretized space, the Green’s function is finite and diverges as
(resolution)3 in 3d. To avoid this singularity, we need to choose window functions which decay
faster than |ω |3 at large |ω |. Two natural candidates are the difference of two Lorentzians [51]
and the square of a Lorentzian.
5.2. Square of a Lorentzian

To ensure that the W (ω ) decays faster than 1/|ω |3, we propose the window function,
2Γ̃3 /π
W (ω ) = 2 , (20)
(ω − ω̃ )2 + Γ̃2
which is a normalized square of a Lorentzian function. This window function has a double pole
at ω = ω̃ + iΓ̃ in the upper-half plane. Applying the residue theorem, we have from Eq. (16)
and Eq. (20)
Z ∞
LDOS(ω )W (ω )dω ≈ Re f (ω̃ + iΓ̃) − iΓ̃ f ′ (ω̃ + iΓ̃) ,

L= (21)
−∞
where f ′ (·) denotes differentiation with respect to ω . In appendix 10, we show that
f (ω , x′ ) 12
Z
f ′ (ω , x′ ) = +i ε (x)ET (x, ω )E(x, ω )dx. (22)
ω π
From Eq. (12) and Eq. (22), it is clear that both f (ω̃ + iΓ̃) and f ′ (ω̃ + iΓ̃) can be obtained from
a single scattering solution E(x, ω̃ + iΓ̃) (see appendix 10) and
f (ω̃ + iΓ̃) − iΓ̃ f ′ (ω̃ + iΓ̃)

ω̃ 6 12
Z
= (− )ê∗j · E(x, ω̃ + iΓ̃) + Γ̃ ε (x)ET (x, ω̃ + iΓ̃)E(x, ω̃ + iΓ̃)dx. (23)
ω̃ + iΓ̃ π π
In summary, we can still obtain the entire frequency-averaged LDOS by solving a single scat-
tering problem Eq. (17) at a complex frequency ω̃ + iΓ̃.
We know that Eq. (20) gives a finite average LDOS because it decays fast enough with ω ,
but it is interesting to also consider how it fixes the divergence from the second viewpoint in
Sec. 5.1 (that of the infinite power absorption from a dipole in a lossy medium). The explanation
is essentially that the second term in Eq. (23) is roughly a subtraction of the divergent absorbed
power from the first term: Γ̃ε is ω Im(ε̃ ) from Sec. 5.1, and ω Im(ε̃ )|E|2 is absorbed power [41].
(A subtlety arises from having +ET E, rather than −|E|2 , but the 1/r3 divergence at r → 0
should be dominant in Im(E) for small Γ̃ so one should have ET E ≈ − Im(ET ) Im(E) ≈ −|E|2
as r → 0.)
Since the role of the second term in Eq. (23) is essentially to subtract off the divergent ab-
sorbed power in lossy ε̃ medium, and this divergence comes from the 1/r3 field divergence
that is independent of geometry (the scattered field from the surrounding geometry is finite at
r = 0), one might expect that the second term in Eq. (23) plays little role in geometry optimiza-
tion at a fixed resolution. Indeed, we find in numerical experiments that the optimizations with
and without the second terms in Eq. (23) for the 2d TE case (discussed in Sec. 8.2) discover
similar structures. Therefore, in Sec. 8 we optimize the simpler single-Lorentzian objective of
section 5.1, although Eq. (23) is computationally feasible if it becomes necessary.
6. A preliminary formulation
Now we have a preliminary formulation for our cavity optimization in terms of the frequency-
averaged local density of states:
Z ∞
max L = LDOS(ω )W (ω )dω . (24)
{designs} −∞
We can evaluate the objective L by contour integration, which only requires us to solve the com-
plex scattering problem Eq. (17) once. If we choose the window function W (ω ) from Eq. (20),
then the problem can be reformulated as
max L = Re L = f (ω̃ + iΓ̃) − iΓ̃ f ′ (ω̃ + iΓ̃) .

(25)
{designs}
Mathematically, to compute the objective in Eq. (25), we need

1. For given ε (x), ω̃ , and Γ̃, solve the complex scattering problem Eq. (17) to obtain
E(x, ω̃ + iΓ̃).
2. Obtain L from the solution E(x, ω̃ + iΓ̃) through the formula Eq. (23).
3. Take the real part of L to get L.
To speed up the optimization, we must also have the gradient of the objective with respect
to the design parameter εk (the dielectric constant at each “pixel” k). Applying the standard
adjoint technique [54], only one more linear system with the same operator M˜(ε , ω̃ ) but a
different source term need be solved to obtain the gradient. We provide the detailed calculation
in appendix 10 and summarize the procedure here:
1. Solve the complex scattering problem
M˜(ε , ω̃ )A(x, ω̃ + iΓ̃) = ε (x)E(x, ω̃ + iΓ̃) (26)
to obtain A(x, ω̃ + iΓ̃).
2. The gradient ∂ L /∂ εk is a combination of E(x, ω̃ + iΓ̃) and A(x, ω̃ + iΓ̃) (see appendix
10):
∂L

1 2
= i+ ω̃ ET (xk , ω̃ + iΓ̃)E(xk , ω̃ + iΓ̃)
∂ εk Q̃ π
12 ω̃ 3

i
+ 1+ AT (xk , ω̃ + iΓ̃)E(xk , ω̃ + iΓ̃). (27)
π Q̃ 2Q̃
3. Take the real part of ∂ L /∂ εk to obtain ∂ L/∂ εk .

Note that the scattering operator Eq. (26) in the sensitivity analysis is the same as the operator
in the objective evaluation. We can take advantage of this by reusing the information (e.g., the
preconditioner or LU factorization) from the solution of Eq. (17). We will discuss this in detail
in Sec. 7.1.
7. Numerical scheme for cavity optimization

In this section, we discuss the numerical implementation for our frequency-averaged LDOS
formulation given in Sec. 6. In order to solve this PDE-constrained optimization problem com-
putationally, we need fast and efficient implementations for objective evaluation, gradient eval-
uation, and optimization.
7.1. Objective and gradient evaluation

As we discussed in Sec. 6, evaluating the objective LDOS involves solving the scattering prob-
lem Eq. (17). We can apply any standard frequency-domain solver technique to this problem
(e.g., finite-difference, finite-element, or boundary-element methods). Here, we simply adopt
the finite-difference approach [55–57]. If we impose mirror symmetry planes in the system, we
can obtain an 8-fold reduction in the number of unknowns (see Sec 8.5.1).
For the finite-difference frequency-domain (FDFD) method, the most robust solution tech-
nique is a sparse-direct solver, which is excellent in 2d, but expensive (in both memory and
time) in 3d [58]. In contrast to direct solvers, iterative solvers (e.g., GMRES or BiCGStab) work
quite well if one has a good preconditioner [36]. Here, we combine both of these techniques.
During the optimization, we re-solve many times for slightly different structures. Therefore,
we can use sparse-direct factorization from one step as a preconditioner for iterative solvers in
many subsequent steps [51]. We implemented the FDFD solver with the parallel sparse-matrix
library PETSc [59–61] and the parallel sparse-direct solver PaStiX [62].
7.2. Optimization scheme

We use a free-software implementation [63] of standard gradient-based optimization algo-
rithms, and we find that low-storage BFGS [64] and conservative convex separable approxi-
mations (CCSA) [65] work equally well. However, we find that one additional mathematical
transformation is required in order to optimize high-Q cavities: instead of maximizing L, we
minimize 1/L.
If we attempt to maximize L directly as in Eq. (24), we typically find that the convergence
of any standard optimization algorithm slows to a halt for Q & 1000. The reason for this is
that, for high-Q cavities, the objective function becomes a “narrow ridge” (a sharply-peaked
function along some low-dimensional manifold in the parameter space), with a large second
derivative (∼ Q3 , as shown in [51]) in the direction perpendicular to the ridge, and it is well
known that optimization along narrow ridges is problematic [66]. Most algorithms tend to “zig-
zag” slowly along the ridge, and even quasi-Newton methods like BFGS break down when the
second derivative is so large that the Hessian matrix becomes ill-conditioned. However, because
the LDOS is strictly positive, there is a simple solution: maximizing L is equivalent to minimiz-
ing 1/L, and the reciprocal of a sharp peak is a shallow valley, so we find that the 1/L objective
avoids most of the problems of slow convergence. (For example, in the 2d TM optimization to
be discussed in Sec. 8.1, maximization of L makes no progress if the initial structure is a high-
Q photonic-crystal cavity, whereas minimization of 1/L converges to a significantly improved
structure as described below.)
In practice, there is another useful technique: a successive-refinement strategy (somewhat
analogous to [67,68]). We found that gradually increasing the specified Q̃ (decreasing the band-
width 1/Q̃) tends to reduce the likelihood that the optimization becomes trapped in a poor local
optimum close to the starting structure, and usually produces a much better result at the highest
Q̃. That is, we optimize for Q̃ = 10, then use that result as a starting point for optimization at
Q̃ = 100, and so on.
8. Results for cavity optimization

In this section, we present some 2d and 3d cavity-optimization results, summarized as fol-
lows. We start with high-resolution 2d cases, and run simulations with different initial guesses
(vacuum, photonic crystal with a defect, and random structures) and different dipole polariza-
tions (TM, TE, and random). In the region to be optimized, we allow the dielectric constant
ε ∈ [1, 12.4] at each pixel to be one degree of freedom [Fig. 3(a)]. For the 2d TE case, the
optimization discovers similar structures for maximizing the spontaneous emission rate of a
specific dipole polarization and a randomly polarized dipole. However, optimizing the worst
case of a randomly polarized dipole finds a three-fold symmetric structure (Sec. 8.2.3), while
two-frequency cavity optimizations yield more complex patterns depending on the polarization
(Sec. 8.3). In another 2d scenario, to obtain the Q versus V tradeoff analogous to 3d slabs,
we limit the degrees of freedom in one direction and choose a thin strip, instead of a square,
as the region for optimization [Fig. 3(b)]. As the degrees of freedom increase, the radiation Q
first increases and then becomes saturated, limited by the numerical precision in the computa-
tion. Finally, we ran a full 3d optimization on a supercomputer
√ and obtained a structure with
extremely small mode volume V = 0.06(λ /n)3, (n = 12.4). In the following, we are mini-
mizing 1/L (the inverse of the frequency-averaged LDOS) and variations thereof, as described
in the previous sections, but for simplicity we describe this below as “maximizing LDOS.”
Here, the focus is on illustrating the formulation and on establishing theoretical bounds. Fur-
ther regularization techniques are generally required to obtain easily manufacturable structures,
as discussed in Sec. 9.
8.1. 2D TM case
In this section, we maximize the LDOS for a dipole with TM polarization (E out of plane)
in a 2d setting [Fig. 3(a)]. One possible starting point is a photonic crystal with a defect [14],
like the one shown in Fig. 4(a). This is a periodic arrangement (periodicity a) of dielectric
silicon rods (radius 0.2a and permittivity ε = 12.4) with one defect rod at the center (radius
0.1a). The defect TM mode is at frequency 0.32(2π /a), with quality factor Q = 1.41 × 108 and
mode volume V = 0.097(λ /n)2. With this structure as an initial guess, we run the optimization
and obtain an entirely different nested-ring structure [Fig. 4(b)] with quality factor Q = 1.01 ×
1010 and mode volume V = 0.075(λ /n)2. Clearly, the optimization itself discovers a “radially
periodic” structure, reminiscent of a Bragg onion [69] . We also run the optimization with
vacuum as initial guess and obtain similar structure (Fig. 5) with Q = 1.30 × 109 and V =
0.075(λ /n)2.
In these two optimizations, we gradually increase the specified Q̃ (we decrease the bandwidth
(a). A square region for degrees of freedom. (b). A thin strip region for degrees of freedom.
Fig. 3. For 2d cavity optimization, we start in Secs. 8.1–8.3 by optimizing over every pixel
in the interior of the computational domain as indicated in (a). This leads to cavities that
utilize bandgap structures to confine light with arbitrary Q, regardless of V , limited only
by the size of the domain. In order to investigate Q vs. V tradeoffs analogous to those in
3d slabs, in Sec. 8.4 we limit the degrees of freedom to a thin strip (b), which imposes
intrinsic radiation losses (perpendicular to the strip) and forces the optimization to sacrifice
V in order to increase Q. A full 3d optimization is considered in Sec. 8.5.
1/Q̃) from 10 to 105 . The optimization at Q̃ = 10 actually gives a high Q cavity (almost the same
radiation Q) with the resonance at about 1.003ω̃ . The optimizations at higher Q̃ simply tune
this structure so that the resonant frequency becomes much closer to ω̃ . In a two-dimensional
situation such as this, there is no intrinsic Q versus V tradeoff, unlike in 3d slabs [14], because
Q → ∞ for a finite modal volume V as the number of layers in a Bragg onion increases [69].
So, the optimization in such a 2d case is mainly minimizing V , while the Q is bounded only by
the computational-cell size.
Note that we allow the dielectric permittivity of each pixel to vary continuously from
εmin = 1.0 to εmax = 12.4, but almost all the pixels (except a few at the interfaces) are at either
εmin or εmax in the optimized structure. This phenomenon (reminiscent of “bang-bang” solutions
in control theory [70]) had also been observed empirically in other cavity-related optimization
work [3,71,72]. There has been some recent progress in proving theoretically that this is the ex-
pected solution: Osting and Weinstein [73] recently analyzed optimization problems for scalar
waves, and showed that maximizing an energy confinement time over the permittivity at every
point in space generally leads to a solution in which the permittivity is either the maximum or
the minimum allowed value at every point, excepting a set of measure zero (e.g., at the inter-
faces between regions) in the limit as the resolution goes to infinity. However, we also obtain
counterexamples for other types of cavity optimization, e.g. in the doubly-degenerate cavity
design of Sec. 8.2.3, where substantial regions converge to intermediate values. Fortunately, in
cases where this occurs, there are a variety of techniques to obtain binary solutions as discussed
in Sec. 9.
8.2. 2D TE case
In this section, we consider the 2d TE polarization, for which (unlike the TM polarization
of Sec. 8.1) the objective function breaks rotational symmetry and we do not expect similar
Bragg-onion solutions. We will start with maximizing LDOS for a fixed êx polarization in
(a). PhC cavity initial guess for 2d TM optimization. It has quality factor
Q=1.41×108 and mode volume V = 0.097(λ /n)2 .
(b). 2d TM optimized structure with Q=1.01×1010 and V = 0.075(λ /n)2 obtained from PhC initial
guess.
Fig. 4. 2d TM optimization from PhC cavity initial guess (Sec. 8.1).
Fig. 5. 2d TM optimization from vacuum initial guess (Sec. 8.1). Q=1.30×109 and V =
0.075(λ /n)2 .
8.2.1. For a randomly polarized dipole, we consider two cases: maximizing the average over all
polarizations in 8.2.2 and maximizing the minimum (worst case) over all polarizations in 8.2.3.
8.2.1. Fixed dipole polarization: max LDOS(ω̃ ; êx )

For the 2d TE polarization, let us first look at the case where the dipole is polarized in the êx
direction. In other words, we want to maximize LDOS(ω̃ ; êx ). From a vacuum initial guess,
the optimization discovers the structure shown in Fig. 6. This structure has quality factor
Q=5.16×108 and mode volume V = 0.092(λ /n)2. [Again, the Q̃ = 10 gives an equally high-Q
cavity with resonant frequency at 1.0007ω̃ . The optimizations at higher Q̃ = 102 to 105 simply
tune the resonant frequency to ω̃ .] For a random initial guess (uniform random pixels), we also
obtain the same structure as the one from a vacuum initial guess, which suggests that the result
may be close to a global optimum.
8.2.2. Randomly polarized dipole: max mean j LDOS(ω̃ ; ê j )

Now we want to study the case in which the dipole is randomly polarized in the plane and the
objective is the mean LDOS. One might hope that the optimization will find a structure that
resonates for both polarizations at the same frequency, and hence by linearity will resonate for
all in-plane polarizations—this corresponds to the requirement that the microcavity be doubly
degenerate, and many symmetry groups besides circular symmetry can support double degen-
eracies (such as three-fold, four-fold, or six-fold symmetrical structures [74]). However, we find
that this is not the case: the optimization finds a singly resonant cavity that enhances LDOS for
one polarization (chosen “randomly” depending on the initial structure) at the expense of the
other polarization. As explained below, this suggests that the LDOS of the best single-mode TE
cavity is more than twice the LDOS of the best doubly degenerate TE cavity. We performed
the mean-LDOS optimization as follows. It is easy to show [29, problem 8.6] that maximizing
the LDOS for a random polarization by averaging all polarizations is equivalent to maximizing
Fig. 6. 2d TE optimization for êx polarization from Sec. 8.2.1. The structure has
Q=5.16×108 and V = 0.092(λ /n)2 obtained from vacuum initial guess.
the mean from the êx and êy polarizations, namely [LDOS(ω ; êx ) + LDOS(ω ; êy )] /2. For this
new objective, we ran many different simulations with different pseudo random initial guesses
(each pixel is randomly chosen between εmin and εmax ). About half gave the structure optimiz-
ing the êx polarization (Fig. 6), while the other half gave the 90◦ -rotated structure optimizing
the êy polarization. From these results, it seems that the optimization, instead of favoring both
êx and êy polarization simultaneously, simply randomly picks one direction and optimizes it.
That is, there is a spontaneous symmetry breaking: it is better to optimize one polarization at
the expense of the other than to try to obtain a doubly degenerate cavity that resonates for both
polarizations. By selecting only one polarization to enhance, these structures sacrifice a factor
of 2 in the mean LDOS, which is why we conclude that the best single-mode LDOS is at least
twice as big as the best two-mode LDOS.
8.2.3. Optimization for a randomly polarized dipole: max min j LDOS(ω̃ ; ê j )

In the previous section, we showed that maximizing the mean LDOS over all in-plane polariza-
tions was equivalent to maximizing the LDOS for a single polarization at the expense of the or-
thogonal polarization. In order to obtain a structure that enhances all polarizations equally (via
a doubly degenerate resonance), we consider a different objective: we maximize the minimum
LDOS over all in-plane polarizations (rather than the mean). The result of this optimization is
a three-fold symmetric microcavity shown in Fig. 7(a), which is the smallest symmetry group
that supports a (non-accidental) doubly degenerate mode [74] shown in Fig. 7(b-c). (The rectan-
gular FDFD grid breaks the three-fold symmetry, but the structure converges to true three-fold
symmetry with an exact degeneracy as the resolution is increased.) Intuitively, larger symmetry
groups have fewer degrees of freedom for optimization (the Cmv symmetry group [74] with
m-fold rotational symmetry plus reflections leaves only a wedge of angle π /m in the degrees
of freedom), so the optimization is picking the smallest possible symmetry group in order to
maximize the number of degrees of freedom available for maximizing LDOS.
As predicted in Sec. 8.2.2, the effective modal volume V of this doubly degenerate cavity is
0.34(λ /n)2 > 2 times the volume of the single-mode cavity from Sec. 8.2.1.
Technically, maximizing the minimum LDOS over all polarizations is significantly harder
than maximizing the mean LDOS. Unlike the mean LDOS, it is not sufficient to consider only
the êx and êy polarizations, and in principle we must solve for the LDOS at all angles. The
simplest approach is to merely sample the set of polarizations to compute the LDOS at many
discrete angles, and then to maximize the minimum LDOS) over this set. Although we tried
sampling up to 20 discrete angles, we found that it was sufficient (obtaining the same structure)
to sample only three angles (0, 60, and 120). Furthermore, the minimum LDOS over several
angles is no longer a differentiable function, but we circumvent that problem by the standard
transformation [75] of introducing a “dummy” variable t and solving maxt subject to the non-
linear constraints t ≤ LDOS j at each angle j. (The resulting nonlinear-programming problem
was solved via the CCSA algorithm [65].)
8.3. 2D optimization for different frequencies: max minn LDOS(ω̃n ; ê j )

In nonlinear devices, e.g. for nonlinear frequency conversion, it is often desirable to design a
cavity which resonates at multiple distinct frequencies [9, 10, 76–79], leading to a challeng-
ing multi-frequency cavity-design problem if one desires a small-volume cavity [80, 81]. The
simplest approach to designing such a multi-frequency cavity with our optimization techniques
is to maximize the minimum LDOS for sources at two (or more) frequencies, ω̃1 and ω̃2 . For
example, we begin by considering ω̃2 = 2ω̃1 (coupling TM to TM or TE to TE), the desired
relationship for intra-cavity second-harmonic generation (SHG) [79, 81], with results shown
in Figs. 8(a) and 8(c). These results exhibit Bragg-onion-like structures similar to the single-
frequency optimization in the previous sections, which is not surprising considering that Bragg
mirrors tend to have gaps at integer frequency multiples [14] and so the same mirror can con-
fine both a fundamental and harmonic frequency. A more challenging case for optimization,
therefore, is to design frequencies that are not integer multiples, and for illustration purposes
we considered ω̃2 = 1.5ω̃1 , which results in the more complicated structure shown in Figs. 8(b)
and 8(d).
8.4. 2D TE thin strip case

In previous 2d TM and TE polarization cases, we expect and obtain almost no Q versus V trade
off since the cavity can be surrounded by complete photonic bandgap or a Bragg onion, with
Q only limited by the computational-cell size. In a 2d setting, to get a Q versus V trade off
analogous to 3d slabs, we need to limit the degrees of freedom in one direction in order to force
the possibility of radiation loss. In this section, we choose the region for degrees of freedom to
be a thin strip [Fig. 3(b)].
Although we expect cavities in a finite-thickness strip to have intrinsic radiation loss in the
direction transverse to the strip, we also expect that it should be possible to make this radiation
loss arbitrarily small at the expense of modal volume. For example, if one starts with a periodic
waveguide structure [14] and introduces a defect in the periodicity, a resonant mode can be
trapped, and the Q can be made arbitrarily large by gradually tapering from the defect to the pe-
riodic structure [14,82–84]. To explore this tradeoff, we limit the modal volume by considering
a finite-length d × 1λ strip of degrees of freedom. For each value of d (d = 1, 2, 3, 5λ ), we plot
the obtained radiation lifetime Qrad as a function of the requested bandwidth Q̃, and the result is
shown in Fig. 9(a). For any given d, as Q̃ is increased then Qrad is increased as well (as explained
in Sec. 4), until it saturates at a maximum determined by two factors. First, the maximum Qrad
is limited by d: a longer d gives more degrees of freedom and increases the maximum possible
(a). Optimized structure for randomly polarized dipole.
(b). Hz field of first degenerate mode. (c). Hz field of second degenerate mode.
Fig. 7. Optimized doubly-degenerate cavity from Sec. 8.2.3, generated by maximiz-

ing the minimum LDOS over all in-plane polarizations. Starting from a vacuum initial
guess, the optimization discovers a structure with C3v (3-fold) rotational symmetry (top),
which supports doubly-degenerate [74] TE modes whose Hz fields are plotted at bottom
(blue/white/red = negative/zero/positive). (The rectangular FDFD grid slightly breaks the
three-fold symmetry and the degeneracy, but the results converge to exact C3v symmetry
and degeneracy as resolution increases.)
(a). Optimization for TM with ω̃2 = 2ω̃1 . (b). Optimization for TM with ω̃2 = 1.5ω̃1 .
(c). Optimization for TE with ω̃2 = 2ω̃1 . (d). Optimization for TE with ω̃2 = 1.5ω̃1 .
Fig. 8. Two-frequency cavity optimization from Sec. 8.3, for either TM (top) or TE (bottom)
polarizations. Left: microcavities which maximize the minimum LDOS at two frequencies
ω̃1 and ω̃2 = 2ω̃1 , e.g. for intra-cavity second-harmonic generation applications [79, 81].
Confinement of such integer-multiple frequencies is physically enabled by the fact the con-
centric Bragg-onion “1d” bandgaps tend to occur at integer-multiple frequencies [14]. A
more challenging case is a two-frequency cavity for ω̃2 = 1.5ω̃1 , resulting in the more
complicated structures shown at right. (All structures were optimized from vacuum initial
guesses.)
Qrad , as expected. Second, around Qrad ≈ 107 the improvement becomes limited by the numer-
ical precision, which prevents the optimization from making progress even though higher Qrad
should theoretically be possible. For d = 5λ , the resulting structure is shown in Fig. 9(b). [Note
that the data in Fig. 9(a) are from the optimization result for the objective ε (x′ )LDOS(ω , x′ ).]
8
10
DOF:5.0λ*1.0λ
DOF:3.0λ*1.0λ
7 DOF:2.0λ*1.0λ
10
DOF:1.0λ*1.0λ
6
Qrad 10
5
10
4
10
3
10
2
10 1 2 3 4 5 6
10 10 10 10 10 10
Q̃
(a). Qrad vs Q̃ for 2d thin strip with different length (different DOF).
(b). An optimized structure for 2d thin strip from vacuum initial guess.
Fig. 9. 2D TE optimization for thin strips with fixed width (geometry sketched in Fig. 3b).
Fig. (a): Qrad vs. Q̃ for 2d thin stirps with same width (λ ) but different length (d =
λ , 2λ , 3λ , 5λ ). As Q̃ is increased in the optimization, higher Qrad are obtained until Qrad is
limited by the degrees of freedom. As the degrees of freedom increase, Qrad first gets big-
ger, but becomes saturated at some level around 107 due to numerical precision. Fig. (b):
An optimized 2d thin-strip structure with width λ and length d = 5λ .
8.5. 3D case
In this section, we present the optimization results for the TE-like polarization in the 3d slab
setting (Fig. 10). We also briefly discuss some variations on the optimized structure and its
comparison with an air-slot cavity.
8.5.1. 3D slab optimization

With the optimization tools developed in the previous sections, we run large-scale simulations
on 3d slab case (with an in-plane dipole source, coupling to even-symmetry “TE-like” [14] res-
onances). Here we choose the dimensions of the slab to be 3λ -3λ -0.19λ , where the thickness is
0.19λ = 0.67(λ /n). A sketch of the physical model is shown in Fig. 10(a), and the real compu-
tation domain (with mirror-symmetry reductions) is illustrated in Fig. 10(b). The optimization
(a). Sketch of the region for degrees of freedom in 3d slab. (b). Sketch for 3d slab computational do-
main (with mirror-symmetry imposed).
Fig. 10. We wish to optimize a microcavity in an air-membrane Si slab in Sec. 8.5, with
the effective computational domain depicted in (a), where the degrees of freedom are every
pixel in the 2d pattern of the slab cross-section for a fixed thickness. Since all 2d single-
polarization optimizations found structures with two mirror symmetry planes, we can re-
duce the computational domain to 1/8 the volume (b) by imposing these mirror symmetries
along with the vertical mirror symmetry.
discovers a structure (Fig. 11) with quality factor Q = 30000 and extremely small mode vol-
ume V = 0.06(λ /n)3. [This result is obtained from optimizations with absorption Q̃ gradually
increasing from 10 to 104 . The optimization discovers structures with radiation Q = 1.18 × 104
at Q̃ = 100, with radiation Q = 2.55 × 104 at Q̃ = 1000, and with radiation Q = 2.98 × 104 at
Q̃ = 104 .]
A comparison with other large- or small-scale optimization work, such as 2.5d optimiza-
tion [72], L3 -type cavity [20] and H0 -type cavity [1] optimization are given in table 1. Clearly,
the optimization was able to achieve four times smaller mode volume than the smallest mode
volume (for Q within one order of magnitude of ours) we found in the literature [1]. (Note that
L3 -type [20] and H0 -type [1] cavities are designed from small-scale optimization and can be
fabricated; while both 2.5d optimization [72] and our results are purely theoretical and compu-
tational investigation from a large-scale optimization perspective.)
Table 1. Comparison of Q and V for structures from our result and the literature.
Optimization Quality Factor Q Mode Volume V (λ /n)3
2.5d optimization [72] 8000 0.32
L3 -type cavity optimization [20] 100000 0.70
H0 -type cavity optimization [1] 280000 0.23
LDOS optimization 30000 0.06
8.5.2. Sensitivity to small features

In the optimized structure, there are some filament-like structures which would be difficult to
fabricate at infrared scales. As a first assessment of the importance of these tiny features, we
Fig. 11. Optimized pattern for a 3d slab from vacuum initial guess (Sec. 8.5.1) with dimen-
sions 3λ -3λ -0.19λ : Q=30000 and V = 0.06(λ /n)3 .
Fig. 12. 3d slab structure after manually removing tiny features in Fig. 11 (Sec. 8.5.2):
Q = 10000 and V = 0.06(λ /n)3 .
manually remove the filaments and obtain a structure (Fig. 12) with Q = 10000 and roughly
same V . Instead of post-processing the structure in this way, which has the disadvantage of
no longer being a local optimum, future work should consider suppress these tiny features by
imposing some explicit constraints during optimization or by some regularization and projec-
tion [3] as further discussed in Sec. 9.
8.5.3. Comparison with air-slot cavity

All these cavities listed in table 1 are dielectric cavities. In other words, the centers of these cavi-
ties are high-dielectric materials (Si and GaAs) and these cavities are useful for dipoles/emitters
lying in these materials. It is also reflected in our unit of mode volume. For example, the mode
volume of the cavity we obtain is V = 0.06(λ /n)3 = 0.06(λ /nSi)3 .
It is known that air-slot cavities [85–88] can have extremely small volumes. For example,
Nomura [86] reported an air-slot cavity with Q = 4.8 × 106 and V = 0.015(λ /nair)3 . Although
0.015 is smaller than 0.06, these two kinds of cavities are not comparable in two ways. First,
the two mode volumes are in different units (λ /nair )3 versus (λ /nSi )3 . In our units, their modal
volume is 0.65(λ /nSi)3 . Second, these two types of cavities are for different applications: air-
slot cavities are useful for emitters lying in air, while the semiconductor-based cavities are
designed for emitters lying in Si and GaAs.
If the application is for emitters lying in air, in theory, we can also introduce an infinitesimal
air-slot at the center, oriented perpendicular to the electric field, into our structure. As discussed
in [85], after the introduction of an air-slot, the mode volume decreases by a factor of (nSi /nair )2
without changing Q. In our case, this factor is about 12.4, and the new mode volume is 4.8 ×
10−3 (λ /nSi )3 = 1.1 × 10−4(λ /nair )3 . Because the resolution we used (46-pixel per wavelength
in air) is not that high, the optimization discovers a dielectric cavity, instead of one with air-
slot type. In future work, one could run the optimizations with high resolutions (at least in 2d
cases) to investigate whether air-slot structures can be discovered; at such high resolution, the
regularized figure of merit of Eq. (21) discussed in Sec. 5.2 becomes essential.
9. Manufacturability: Projection and Regularization

The primary purpose of this paper is to improve the mathematical formulation of the mi-
crocavity problem to make it amenable to large-scale topology optimization. It is well known,
however, that such topology optimization can sometimes lead to non-manufacturable designs,
due to three problems: regions of intermediate ε values, fine features (such as the “hairs” in
Fig. 11), and extreme sensitivity to variations in the design parameters. Fortunately, there are
a variety of techniques to correct these difficulties, which could easily be combined with our
LDOS formulation, and we briefly review these projection, regularization and robustification
techniques here for the benefit of future work.
In cases where the optimization does not lead to “bang-bang” solutions, i.e. where there are
large regions of intermediate values that do not correspond to available materials, one can use a
variety of penalization techniques that add a penalty (increased as needed) to the objective for
intermediate ε values [3, 89]. Another possibility is to use a level-set method [4, 90, 91], which
guarantee binary solutions (except possibly on a set of measure zero, since intermediate values
are typically used at the level set boundary in order to ensure continuity). A third possibility is
the SIMP (Solid Isotropic Material with Penalization) [3], a re-parameterization of the problem
that is used in conjunction with the filtering techniques described below in order to project the
solution towards a bang-bang result.
To eliminate small features, a common solution is to apply a smoothing filter to the degrees
of freedom (combined with a smoothed Heaviside projection to transform back towards
bang-bang designs) [2, 5–8, 92–95]. This is often viewed as a regularization to ensure that the
problem is well-posed in the sense that the optimum should converge with resolution (rather
than yielding finer and finer features) [3]. Another approach that can eliminate small features
as well as designs with extreme parameter sensitivity is “robust” optimization, in which one
typically optimizes a worst-case design over a set of parameter uncertainties (e.g. a small
“noise” added to each pixel) [67, 96, 97], which has been shown in some cases to also eliminate
small features in topology optimization for photonics [68, 98, 99].
10. Conclusion
In this paper, we presented a novel formulation for large-scale optimization of optical cavities
via frequency-averaged LDOS. With this formulation, we obtained various 2d and 3d cavity-
optimization results for different applications. Our results show that several times smaller modal
volume is theoretically possible for silicon cavities compared to previous work, without sacri-
ficing Q. Many other possibilities present themselves for future work. First, the slab thickness
in the 3d optimization (Sec. 8.5) is fixed, but one can further extend our approach by allowing
the slab thickness to be one additional degree of freedom to improve the figure of merit. Second,
although we focused here on silicon-based cavities, one can easily apply our approach to other
materials (such as lower-index diamond for visible frequencies [100,101]). Since silicon cavity
design has been so heavily studied and many adequate designs are already known, it is espe-
cially in the context of new materials systems that optimization can be valuable. Third, instead
of maximizing LDOS, one can maximize the frequency-averaged total DOS over the whole
computational cell, which is known to be related to the bounds of the light trapping in photo-
voltaic problems [102, 103]. Fourth, one can extend our work on multi-frequency optimization
(Sec. 8.3) by defining a more specialized figure of merit for nonlinear frequency conversion
that includes an overlap factor [79, 81] for the two modes.
Appendix A: Computation of f ′ (ω , x′ )
In this section, we compute f ′ (ω , x′ ), the differentiation ∂ /∂ ω of f (ω , x′ ) from Eq. (12), to
show that the residue of Eq. (21) can be evaluated without any additional matrix solves. Differ-
entiating on both sides of Eq. (8), we have
∂ E(x, ω ) ∂ M (ε , ω )
M (ε , ω ) + E(x, ω ) = iJ(x)
∂ω ∂ω
∂ E(x, ω ) ∂ M (ε , ω )

=⇒ = M −1 iJ(x) − E(x, ω ) = M −1 (iJ(x) + 2ωε (x)E(x, ω )).
∂ω ∂ω
(28)
Now differentiating on Eq. (12), we have
6 ∂ E(x, ω )
Z
f ′ (ω , x′ ) = − J∗ (x) dx
π ∂ω
6 1 ∗
Z Z
=− J (x)M −1 (iω J(x))dx + 2ω J∗ (x)M −1 ε (x)E(x, ω )dx
π ω
(29)
6 1 ∗ 12
Z Z T
=− J (x)E(x, ω )dx + i M −1 (iω J(x)) ε (x)E(x, ω )dx
π ω π
f (ω , x′ ) 12
Z
= +i ε (x)ET (x, ω )E(x, ω )dx.
ω π
Note that we use the properties that M (ε , ω ) is complex symmetric (M = M T ) for both real
and complex ω and J(x) is real [J∗ (x) = JT (x)].
Appendix B: Computation of the objective and its gradient
In this section, we derive an efficient expression for the objective L =
Re f (ω̃ + iΓ̃) − iΓ̃ f ′ (ω̃ + iΓ̃) defined in Sec. 5.2 as well as its gradient. The gradient

of the objective with respect to the design parameters is calculated with standard adjoint
methods [54] in order to minimize the number of required matrix solves.
Let us denote its complex version by L = f (ω̃ + iΓ̃) − iΓ̃ f ′ (ω̃ + iΓ̃). From Eq. (12) and
Eq. (22), we can simplify it as
L = f (ω̃ + iΓ̃) − iΓ̃ f ′ (ω̃ + iΓ̃)

f (ω̃ + iΓ̃, x′ )

12
Z
= f (ω̃ + iΓ̃) − iΓ̃ +i ε (x)E (x, ω̃ + iΓ̃)E(x, ω̃ + iΓ̃)dx
T
ω̃ + iΓ̃ π
ω̃ 12 (30)
Z
= f (ω̃ + iΓ̃) + Γ̃ ε (x)ET (x, ω̃ + iΓ̃)E(x, ω̃ + iΓ̃)dx.
ω̃ + iΓ̃ π
ω̃ 6 ∗ 12
Z
= (− )ê j · E(x, ω̃ + iΓ̃) + Γ̃ ε (x)ET (x, ω̃ + iΓ̃)E(x, ω̃ + iΓ̃)dx.
ω̃ + iΓ̃ π π
In the rest of this section, we compute the gradient of L with respect to the design parameter
εk , which is the dielectric constant at x = xk . To obtain the sensitivity of E(x, ω̃ + iΓ̃) to εk , we
differentiate Eq. (17) with respect to εk
∂ E(x, ω̃ + iΓ̃) ∂ M˜(ε , ω̃ )

M˜(ε , ω̃ ) + E(x, ω̃ + iΓ̃) = 0
∂ εk ∂ εk
(31)
∂ E(x, ω̃ + iΓ̃)

i
=⇒ =M ˜ −1
ω̃ 1 +
2
δ (x − xk )E(x, ω̃ + iΓ̃) .
∂ εk 2Q̃
Therefore, from Eq. (12), Eq. (22) and Eq. (31), we have
∂ J∗ (x)E(x, ω̃ + iΓ̃)dx ∂ E(x, ω̃ + iΓ̃)

R Z
= J∗ (x) dx
∂ εk ∂ε
Zk
i
= ω̃ 2 1 + (M˜−1 J(x))T δ (x − xk )E(x, ω̃ + iΓ̃)dx
2Q̃
Z (32)
i
= −iω̃ 1 + ET (x, ω̃ + iΓ̃)δ (x − xk )E(x, ω̃ + iΓ̃)dx
2Q̃

i
= −iω̃ 1 + ET (xk , ω̃ + iΓ̃)E(xk , ω̃ + iΓ̃),
2Q̃
and
∂ ε (x)ET Edx ∂ E(x, ω̃ + iΓ̃)
R Z Z
=2 ε (x)ET dx + δ (x − xk )ET (x, ω̃ + iΓ̃)E(x, ω̃ + iΓ̃)dx
∂ εk ∂ εk

i
Z
= 2 ε (x)ET M˜−1 ω̃ 2 1 + δ (x − xk )E(x, ω̃ + iΓ̃)
2Q̃
(33)
+ ET (xk , ω̃ + iΓ̃)E(xk , ω̃ + iΓ̃)
Z
i T
= 2ω̃ 2 M˜−1 ε (x)E(x, ω̃ + iΓ̃) δ (x − xk )E(x, ω̃ + iΓ̃)dx

1+
2Q̃
+ ET (xk , ω̃ + iΓ̃)E(xk , ω̃ + iΓ̃).
Combining Eq. (30), Eq. (32) and Eq. (33), we have
∂L

1 6
= i+ ω̃ ET (xk , ω̃ + iΓ̃)E(xk , ω̃ + iΓ̃)
∂ εk Q̃ π
12 ω̃ 3

i
+ 1+ AT (xk , ω̃ + iΓ̃)E(xk , ω̃ + iΓ̃), (34)
π Q̃ 2Q̃
where A(x, ω̃ + iΓ̃) satisfies the scattering equation
M˜(ε , ω̃ )A(x, ω̃ + iΓ̃) = ε (x)E(x, ω̃ + iΓ̃). (35)
Therefore, as usual for adjoint-methods [54], the gradient of L with respect to design parame-
ters ε can be evaluated with only a single additional matrix solve Eq. (35).
Acknowledgments
X.L. thanks J. Joannopoulos and B. Zhen for helpful discussions. This work was supported
in part by the AFOSR MURI for Complex and Robust On-chip Nanophotonics (Dr. Gernot
Pomrenke) grant number FA9550-09-1-0704 and by the Institute for Soldier Nanotechnolo-
gies under Contract W911NF-07-D0004. This work used the Extreme Science and Engineering
Discovery Environment (XSEDE), which is supported by National Science Foundation grant
number OCI-1053575.

Liang - 2013 - Formulation For Scalable Optimization of Microcavities Via The

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Liang - 2013 - Formulation For Scalable Optimization of Microcavities Via The

Uploaded by

Copyright:

Available Formats

Formulation for scalable optimization of microcavities

via the frequency-averaged local density of states

Citation Liang, Xiangdong, and Steven G. Johnson. “Formulation for scalable

Publisher Optical Society of America

Version Final published version

Citable link http://hdl.handle.net/1721.1/82939

Terms of Use Article is made available in accordance with the publisher's

Xiangdong Liang∗ and Steven G. Johnson

Abstract: We present a technique for large-scale optimization of optical

References and links

Here λ is the vacuum wavelength and n is the index of refraction.

Depending on the applications for enhancement or inhibition, we should maximize or minimize

4.1. Causality and analyticity

4.2. Contour integration

4.3. From complex frequency to material absorption

We denote this complex scattering operator by M˜(ε , ω ), namely

5. Possible window functions

5.1. A simple Lorentzian

5.2. Square of a Lorentzian

f (ω̃ + iΓ̃) − iΓ̃ f ′ (ω̃ + iΓ̃)

max L = Re L = f (ω̃ + iΓ̃) − iΓ̃ f ′ (ω̃ + iΓ̃) .

Mathematically, to compute the objective in Eq. (25), we need

M˜(ε , ω̃ )A(x, ω̃ + iΓ̃) = ε (x)E(x, ω̃ + iΓ̃) (26)

to obtain A(x, ω̃ + iΓ̃).

3. Take the real part of ∂ L /∂ εk to obtain ∂ L/∂ εk .

7. Numerical scheme for cavity optimization

7.1. Objective and gradient evaluation

7.2. Optimization scheme

8. Results for cavity optimization

Fig. 4. 2d TM optimization from PhC cavity initial guess (Sec. 8.1).

8.2.1. Fixed dipole polarization: max LDOS(ω̃ ; êx )

8.2.2. Randomly polarized dipole: max mean j LDOS(ω̃ ; ê j )

8.2.3. Optimization for a randomly polarized dipole: max min j LDOS(ω̃ ; ê j )

8.3. 2D optimization for different frequencies: max minn LDOS(ω̃n ; ê j )

8.4. 2D TE thin strip case

Fig. 7. Optimized doubly-degenerate cavity from Sec. 8.2.3, generated by maximiz-

8.5.1. 3D slab optimization

8.5.2. Sensitivity to small features

8.5.3. Comparison with air-slot cavity

9. Manufacturability: Projection and Regularization

L = f (ω̃ + iΓ̃) − iΓ̃ f ′ (ω̃ + iΓ̃)

∂ E(x, ω̃ + iΓ̃) ∂ M˜(ε , ω̃ )

∂ J∗ (x)E(x, ω̃ + iΓ̃)dx ∂ E(x, ω̃ + iΓ̃)

where A(x, ω̃ + iΓ̃) satisfies the scattering equation

M˜(ε , ω̃ )A(x, ω̃ + iΓ̃) = ε (x)E(x, ω̃ + iΓ̃). (35)

You might also like